The Year Agents Got Real: Claude Opus 4.5 and the 2025 Recap

Nobody has been on my payroll since 2024. That makes me an unusually interested reader of every "agent" announcement, because for me an AI that can really do work is not a novelty. It is capacity.

It is also why I am suspicious of the hype. Yesterday, November 24, Anthropic released Claude Opus 4.5. It closes out a year in which agents stopped being demos and started being products. So here is the honest scorecard: what got real for a small operator in 2025, what is still smoke, and what I would do with it.

What actually changed in the year of agents?

Five releases tell most of the story.

• October 2024: Anthropic introduced computer use, teaching Claude to look at a screen, move a cursor and type. It called the capability experimental and "at times cumbersome and error-prone."³
• July 2025: OpenAI launched ChatGPT agent, which works through a virtual computer to browse, fill out forms and create files, and is trained to ask approval before actions with real-world consequences like a purchase.⁴
• September 2025: Claude Sonnet 4.5 shipped with a 61.4% score on the OSWorld computer-use benchmark and an observation of 30-plus hours of sustained focus on complex tasks.²
• October 2025: Anthropic introduced Agent Skills, and OpenAI introduced apps in ChatGPT.⁵⁶
• November 2025: Claude Opus 4.5 arrived.¹

Be careful with that comparison. The two scores come from different model generations and possibly different test setups, so read it as a direction, not a precise multiple. The direction is a steep climb in under a year.²³

Did agents learn to use a computer?

Yes, with an asterisk the size of a barn.

In 2024, the best system Anthropic tested got 14.9% on OSWorld in the screenshot-only category.³ By September 2025, Sonnet 4.5 scored 61.4%.² That is real progress and also still short of reliable. An agent that gets six of ten computer tasks right is a useful assistant and a dangerous employee.

OpenAI's ChatGPT agent shows the design that works today: it can browse and fill forms, but it is trained to ask permission before a consequential action, and OpenAI calls it experimental.⁴ Anthropic says Opus 4.5 is harder to trick with prompt injection than any other frontier model, though it gives no numeric score.¹ Prompt injection, where a web page hides instructions for the agent, is the risk that keeps me from giving browser agents the keys to anything important.

Capable is not the same as trustworthy. Treat every agent like a new hire on day one: useful, supervised and not yet authorized to wire money.

Why did context and skills matter more than raw intelligence?

Because an agent that forgets is an agent that fails in the middle of the job. (Claude Sonnet 4.5 and the Long-Running Agent goes deeper on this.)

Anthropic's guidance on context engineering describes the real work as curating what the model holds at each step: compaction to summarize and continue, structured note-taking outside the context window, sub-agents for focused tasks, and just-in-time retrieval instead of stuffing everything in the prompt.⁹ Anthropic also says recall accuracy tends to drop as the context window fills, which is why a bigger prompt is not a better prompt.⁹

Agent Skills turned the same idea into a product. Anthropic describes them as folders of instructions, scripts and resources that Claude loads only when needed, usable across Claude apps, Claude Code and the API.⁵ The Model Context Protocol, introduced in November 2024, handles the other half: connecting assistants to the systems where your data lives.¹⁰

Matt Wolfe made the plumbing point earlier in the year, in a video about the Manus agent posted on March 14, 2025: "We didn't need a huge leap in smarter language models to get this agent."¹¹ A small team joined existing tools together, and it felt like a step change. I think that is the pattern for most of 2025.

Plays here, from YouTube

If you take one thing from this section, take this: the know-how you write down is now an asset an agent can use. Your processes, templates and quality checks are worth more than they were a year ago.

What this looks like in a small business

Imagine a weekly report you build by hand every Monday: pull numbers from three tools, check them against last week, write a one-page summary. The old way is a long prompt pasted into a chat. The 2025 way is a written procedure the agent loads when asked for the weekly report, with a template, a list of sources and a short check at the end: do the totals match the source, and are any numbers more than a set amount off last week? The procedure is the product. The model is the engine that runs it.

What does Claude Opus 4.5 add?

According to Anthropic's announcement:¹

• Price: $5 per million input tokens and $25 per million output tokens.
• Efficiency: at medium effort, Opus 4.5 matches Sonnet 4.5's best SWE-bench Verified score while using 76% fewer output tokens. At its highest effort it beats Sonnet 4.5 by 4.3 percentage points while using 48% fewer tokens.
• An effort parameter: developers can trade off speed and cost against capability.
• Product reach: Claude for Chrome available to all Max users, and Claude for Excel in expanded beta.
• Context management: Anthropic reports that context management, memory and subagent coordination together raised a deep research evaluation from 70.48% to 85.30%.

Those are vendor-reported figures on vendor-chosen tests. They tell you the labs are racing on cost and reliability for long tasks. They do not tell you how it will do on your messy folder of client files.

Are businesses actually getting value?

Unevenly, and the data is worth sitting with.

Fortune reported in August on an MIT NANDA study: about 95% of enterprise generative AI pilots showed no measurable profit-and-loss impact, based on 150 leader interviews, a survey of 350 employees and an analysis of 300 public deployments.⁸ The study blamed a "learning gap" in how tools and organizations adapt, not model quality. It also found purchased tools from specialized vendors succeeded about 67% of the time, roughly three times as often as internal builds.⁸ One report is not the final word (why small operators can win), but the pattern matches what I see: tools fail when they are bolted on without a clear job.

Anthropic's own Economic Index offers a view from usage. In its September report, 77% of first-party API transcripts showed automation patterns, compared with 12% showing augmentation, and on Claude.ai automation (49%) overtook augmentation for the first time.⁷ People and companies are giving AI whole tasks, not just asking for help with them.

A year later that sentence is less true, but it is still the right mental model for anything touching money.

What is real, early and hype?

I would put apps in ChatGPT in the "early" box. OpenAI introduced them on October 6 with a handful of launch partners, kept the SDK in preview and said submissions would come later in the year.⁶ The direction is interesting. The practical effect for most small businesses is zero today.

How to judge any agent claim

Ask four questions before you believe a demo.

• What was the test? A benchmark is a score on someone else's tasks. Ask whether those tasks look like yours.
• Who reported it? A vendor's number is a pitch until someone independent repeats it.
• What happens on failure? The best products tell you when they are stuck. The worst keep going and hide the mistake.
• Who is accountable? Software does not carry liability. You do.

What should a small operator do heading into 2026?

Four moves, in order. No heroics.

Three jobs that usually qualify

• Competitor and market research briefs, where every claim has a link you can check.
• Page and funnel audits, where the output is a list of issues against a checklist you wrote.
• First drafts of long documents, where you edit rather than write from a blank page.

Each has a finish line, a check and an undo. None of them sends anything to a customer without you.

The honest summary

Here is the one-sentence version of the year. Agents got real in 2025 where the work is bounded, reviewable and reversible. They stayed unreliable where the work is open-ended, public or financial. That is not a letdown. It is a map, and a good one: it tells you where to spend your first hour and where to keep your hand on the wheel.

The operators who benefit most will not be the ones with the fanciest tools. They will be the ones who can write a clear brief, define a good check and keep a hand on the approval button. That is a management skill, and you can build it now. If you have ever run a team, you already own most of it: scoping work, giving context, reviewing output and giving feedback. If you have not, practice on a small job this week and treat the first result as a draft, not a verdict. The tools will keep changing. The habit of briefing clearly and checking honestly will not. Turning that habit into a working system is what I do in my client builds.

Keep building.Ekai